Advancing the State-of-the-Art in Empirical Privacy Auditing

The paper introduces a novel approach to empirical privacy auditing (EPA) for large language models (LLMs), focusing on the use of synthetic "canary" examples generated through high-temperature sampling (T ≥ 0.8) from LLMs to assess data leakage risks in membership inference and reconstruction attacks. This method enhances the ability to audit models fine-tuned on privacy-sensitive data by ensuring the canaries are high-influence outliers, allowing for effective inspection without compromising real data privacy. The proposed auditing techniques also include a framework for evaluating synthetic data privacy risks, which is critical for practitioners concerned about data leakage in LLM applications.